Deep Latent Competition: Learning to Race Using Visual Control Policies in Latent Space